https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66714
--- Comment #21 from vries at gcc dot gnu.org --- (In reply to cesar from comment #20) > Created attachment 36030 [details] > replace block vars fix > > Tom, thanks for your detailed analysis and reduced test case. As you > suspected, replace_block_vars_by_duplicates isn't updating the > DECL_VALUE_EXPR properly. Yeah, and AFAIU Michaels analysis ( https://gcc.gnu.org/ml/gcc-patches/2015-07/msg01037.html ) he agrees that that is the problem. > That function is setting the value expr to be the > original decl, not the new offloaded copy. My patch teaches it how to use an > offloaded copy. > Cool. > All of the value exprs we're interested in for openacc are INDIRECT_REFs and > I think that holds true for openmp too. Fortran cray pointers caused some > minor problems because those get represented by a INDIRECT_REF to a > CONVERT_EXPR as you in the patch. > > --- a/gcc/tree-cfg.c > +++ a/gcc/tree-cfg.c > @@ -6916,7 +6916,29 @@ replace_block_vars_by_duplicates (tree block, > hash_map<tree, tree> *vars_map, > { > if (TREE_CODE (*tp) == VAR_DECL && DECL_HAS_VALUE_EXPR_P (*tp)) > { > - SET_DECL_VALUE_EXPR (t, DECL_VALUE_EXPR (*tp)); > + tree x = DECL_VALUE_EXPR (*tp); > + > + if (TREE_CODE (x) == INDIRECT_REF) > + { > + tree expr = TREE_OPERAND (x, 0); > + tree decl; > + > + if (CONVERT_EXPR_CODE_P (TREE_CODE (expr))) > + decl = TREE_OPERAND (expr, 0); > + else > + decl = expr; > + > + replace_by_duplicate_decl (&decl, vars_map, to_context); > + > + if (CONVERT_EXPR_CODE_P (TREE_CODE (expr))) > + expr = build1 (TREE_CODE (expr), TREE_TYPE (expr), decl); > + else > + expr = decl; > + > + x = build_simple_mem_ref (expr); > + } How about + else + gcc_unreachable (); ? That makes sure you run into all the unhandled cases. > I tested this patch in gomp-4_0-branch libgomp and everything appears to > work. Does this issue present in trunk too? Yep, this PR was originally filed as gomp-4_0-branch PR but, I changed it to trunk PR at comment 13.
